Semantic Parsing with Dual Learning

This repository contains source code and data for the ACL 2019 Long Paper "Semantic Parsing with Dual Learning".

If you use our framework in your work, please cite it as follows:

    @inproceedings{cao-etal-2019-semantic,
        title = "Semantic Parsing with Dual Learning",
        author = "Cao, Ruisheng  and
          Zhu, Su  and
          Liu, Chen  and
          Li, Jieyu  and
          Yu, Kai",
        booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
        month = jul,
        year = "2019",
        address = "Florence, Italy",
        publisher = "Association for Computational Linguistics",
        url = "https://www.aclweb.org/anthology/P19-1007",
        doi = "10.18653/v1/P19-1007",
        pages = "51--64"
    }

---

Setup

• First, create the environment

    conda create -n sp python=3.6
    source activate sp
    pip3 install -r requirements.txt
  

• Second, pull all the dependencies from remote repository, including evaluator, lib and glove6B word embeddings.

    ./pull_dependency.sh
  

• Construct vocabulary for all datasets in advance under corresponding directory data, run

    python3 utils/statistics.py
  

---

Dataset

---

Experiments are conducted on two semantic parsing dataset ATIS and OVERNIGHT, including traditional train, dev and test files, plus elaborated lexicon files for entity mapping and reverse entity mapping techniques, and extra files for synthesized unlabeled logical forms. An additional ontology file are created for dataset ATIS since there is no evaluator available.

---

ATIS

Files:

• atis_train.tsv: training dataset, 4433 samples.
• atis_dev.tsv: validation dataset, 491 samples.
• atis_test.tsv: test dataset, 448 samples.
• atis_extra.tsv: synthesized logical forms (Lambda Calculus), 3797 samples.
• atis_lexicon.txt: each line specifies a one-to-one mapping between a natural language noun phrase and its corresponding entity representation in knowledge base, such as pair (first class, fist:cl).
• atis_ontology.txt: specify all the entity types, unary, and binary predicates used in the logical form.

Attention: Since there is no evaluator for this domain, we provide a simple type consistency checker for the target logical form (utils/domain/atis_evaluator.py). atis_train.tsv, aits_dev.tsv and atis_test.tsv are preprocessed version provided by Dong and Lapata (2018), where natural language queries are lowercased and stemmed with NLTK, and entity mentions are replaced by numbered markers. For example:

flight from ci0 to ci1	( lambda $0 e ( and ( flight $0 ) ( from $0 ci0 ) ( to $0 ci1 ) ) )

---

OVERNIGHT

It contains eight sub-domains in total, namely basketball, blocks, calendar, housing, publications, recipes, restaurants and socialnetwork.

• [domain]_train.tsv: training and dev dataset. There is no isolate validation dataset in OVERNIGHT. We follow the traditional 80%/20% (train/dev) split in experiments.
• [domain]_test.tsv: test datset.
• [domain]_extra.tsv: synthesized logical forms (Lambda DCS). We revise the template rules in SEMPRE to generate new instances.
• [domain]_lexicon.txt: each line specifies a one-to-one mapping between a natural language noun phrase and its corresponding entity representation in knowledge base, such as pair (kobe bryant, en.player.kobe_bryant ).

Attention: There is also a evaluator program provided by Jia and Liang (2016) in each domain to obtain denotations (utils/domain/domain_overnight.py). Each sample in [domain]_train.tsv and [domain]_test.tsv is of the form:

what player did not play point guard	( call SW.listValue ( call SW.getProperty ( ( lambda s ( call SW.filter ( var s ) ( string position ) ( string ! = ) en.position.point_guard ) ) ( call SW.domain ( string player ) ) ) ( string player ) ) )

---

Experiments

---

Semantic Parsing (Supervised|Pretrain)

Refer to script in run/run_semantic_parsing.sh, for example

./run/run_semantic_parsing.sh dataset_name [attn|attnptr] labeled

dataset_name must be in choices [atis, basketball, blocks, calendar, housing, publications, recipes, restaurants, socialnetwork] and labeled denotes the ratio of labeled examples in training set we are going to use.

---

Question Generation (Supervised|Pretrain)

The procedure is similar to that of Semantic Parsing since we use similar model architecture.

./run/run_question_generation.sh dataset_name [attn|attnptr] labeled

---

Language Model (Unsupervised|Pretrain)

Language model is used to calculate the validity reward during the closed cycle.

./run/run_language_model.sh dataset_name [question|logical_form]

---

Pseudo Method (Semi-supervised)

Use pretrained models of Semantic Parsing and Question Generation to generate pseudo samples. Then shuffle these pseudo samples with labeled samples together to train an improved Semantic Parsing and Question Generation Model.

./run/run_pseudo_method.sh dataset_name [attn|attnptr] labeled

Attention: in the script run/run_pseudo_method.sh, read_sp_model_path and read_qg_model_path are paths to the pretrained models(semantic parsing and question generation). labeled and seed should be kept the same for both the pretraining phases and pseudo method. By default, model type (attn/attnptr) is the same for both semantic parsing and question generation models.

---

Dual Learning (Semi-supervised)

Use pretrained models of semantic parsing, question generation and language models to form two closed cycles with different starting points. Combine dual reinforcement learning algorithm and supervised training together. Running script:

./run/run_dual_learning.sh dataset_name [attn|attnptr] labeled

Attention: in the script run/run_dual_learning.sh, read_sp_model_path, read_qg_model_path, read_qlm_path and read_lflm_path are paths to the pretrained models(semantic parsing, question generation, question language model and logical form language model). labeled and seed should be kept the same for both the pretraining phases and dual learning framework. By default, model type (attn/attnptr) is the same for both semantic parsing and question generation models.